Looped Flows boosts recurrent test-time reasoning
Researchers introduced Looped Flows, a framework that trains recurrent neural network architectures with local denoising objectives to overcome truncated backpropagation limits. By formulating inference as integrating a learned probability flow, the model dynamically scales test-time compute and achieves state-of-the-art results on reasoning benchmarks, including 58.8% accuracy on ARC-AGI-1.
Looped architectures paired with continuous flow formulations represent one of the most promising alternatives to token-heavy autoregressive chain-of-thought, demonstrating that models can scale test-time compute internally without inflating context windows. By decoupling compute from parameter size, looped models spend variable computation on difficult problems through recurrent internal updates rather than bloated token sequences. Structuring recurrence around local denoising objectives with shared noise samples successfully establishes long-range temporal dependencies without full unrolling, solving the historical limits of truncated backpropagation. Furthermore, formulating inference as numerical integration along a probability flow allows dynamic compute tuning at test time, while strong benchmark results like 58.8% on ARC-AGI-1 prove the viability of continuous recurrent reasoning.
DISCOVERED
1h ago
2026-09-12
PUBLISHED
2h ago
2026-09-12
RELEVANCE
AUTHOR
omarsar0